Video Transcript
Phillip Howard:
Anahita and I are AI researchers at Intel Labs, and today we’re going to be showcasing BloomZ running on 8 Habana Gaudi2 accelerators on the Intel Developer Cloud. BloomZ is one of the largest open-source language models available. Today, it has over 176 billion parameters. It was trained to provide short and concise answers to human instructions. So today we’re going to be showcasing it in the context to providing approximately paragraph length answers to questions that we pose.
Anahita Bhiwandiwalla:
To get started, maybe let’s ask BloomZ a generic question on how something works, like how does a telescope work?
Phillip Howard:
Sure. So I’m going to type, “how does a telescope work?” in this terminal window that you see at the top of my screen. And then in the terminal at the bottom, you’re going to see the response from BloomZ along with some diagnostic information about how long the query took, including information that it generated approximately 25 tokens per second.
So here we can see that the response from BloomZ elaborates that a telescope works by magnifying distant objects using lenses and mirrors, and it also mentions that it’s an optical instrument that can be used to see distant objects more clearly than with the naked eye.
Anahita Bhiwandiwalla:
That seems pretty accurate. Maybe let’s ask BloomZ something more technical, like what is K-means algorithm?
Phillip Howard:
Sure. So I’ll ask “what is K-means clustering?”, and after about a second, you’re going to see the response from BloomZ at the bottom saying that it’s an unsupervised learning algorithm that petitions observations into a user-specified number of clusters that each observation belongs to the cluster with the nearest mean.
So I think that’s a pretty accurate description of K-means. Maybe we can ask about another machine learning algorithm, such as “what is a random forest?” So we’re going to see the response from BloomZ is that random forest is an ensemble of decision trees that are trained on different subsets of the data and then use averaging to improve its prediction accuracy and that these can be used for classification or regression problems.
Anahita Bhiwandiwalla:
Those were pretty impressive. This wraps up our live demo of BloomZ running on Habana Gaudi2 accelerators. To use Gaudi2 for your future AI workloads, check out Intel Developer Cloud where it’s available to use today.